Application Of Convolutional Network In Visual Object Tracking

The research of visual object tracking is a significant area in computer vision field,which has widely applications,especially in smart city,automatic driving,intelligent surveillance,intelligent traffic and security.In recent years,the development of machine learning,especially the development of deep learning technology,has made further breakthroughs in the research of object tracking.Based on the study of the feature extraction methods using traditional approaches and deep learning,this thesis researches the application of deep learning technology in visual object tracking.Although deep networks have shown better performance than traditional approaches by offline training,the offline training is hard and time-consuming.In order to solve this problem,this thesis uses a two-layer sparse convolutional network without offline training to extract image feature and provides an online learning approach based on affine propagation clustering method.In order to reduce the search space,the thesis also researches the application of particle filter algorithm on object tracking and use particle filter framework to select candidate target regions.To mitigate the problem of tracking drift in scenarios that target moves fast,a threshold mechanism is introduced to adaptively adjust the parameters in particle motion model,which improves the tracking performance when target moves fast.Combined with the two-layer convolutional network and the particle filter framework,a visual object tracking algorithm is implemented.The system has been tested on 50 challenging videos in a tracking benchmark dataset and compared with other object tracking algorithms.The result shows that the proposed algorithm has good tracking performance and success rate.Good robustness is achieved in complex scenarios with challenges of fast moving,background clutter,occlusion.